• Large-scale 3D Scene Understanding and modeling

    For computer graphics researchers understanding the world from visual data has been a main challenge since the beginning of the community. Research in this area has recently become (again) a focus in the community, leveraging advances in a wide range of topics that include scene parsing, mapping, detection and reconstruction, among others. Nevertheless, there is still a long path until we have more complex approaches capable of a detailed understanding of the environments that surround us. Traditionally, hand-crafted features are widely used in existing methods. However, the recent progress of deep learning techniques show that, the simplest feature like pixels can be directly combined with the neural networks and trained together. We introduce a 3D point cloud labeling scheme based on 3D Convolutional Neural Network.

    Researchers: Qian Xie, Xianglin Guo, Xingyu Xie, Yangxing Sun

  • Automatic 3D Electrical Substation Point Cloud Modeling

    3D modeling of point clouds is an important but time-consuming process, inspiring extensive research in automatic methods. Prior efforts focus on primitive geometry, street structures or indoor objects, but industrial data has rarely been pursued. In this research, we are focusing on modeling electrical substations as a family of industrial infrastructures.

    Researchers: Qiaoyun Wu, Hongbin Yang, Jin Huang

  • Vision based Simultaneously Localization and Mapping

    By observing and exploring surroundings, humans can orient the direction through eyes. The ideas were brought to the robot field, and as computer vision technique improves, vision SLAM was considered the best way to understand and reconstruct the world. Some problems exist in vision SLAM system, like low precision, sparse map, and scene distortion. Deep learning seems have the ability to solve some of them, this could be a new branch to improve the performance of VSLAM system.

    Researchers: Xiaoxi Gong, Dawei Li

  • Autonomous driving system based on deep learning

    Autonomous vehicle could drive safely and efficiently. More importantly, it is the main approach to improve terrible traffic conditions. Though today's researches were still focus on assistance driving system, benefit from big data and deep learning framework, systems become more reliable and intelligent than before. The traditional techniques used in autonomous driving system, such as radar, GPS, odometry, and computer vision. Here we focus on the deep learning method, which can replace expensive laser measuring systems by a set of cameras.

    Researcher: Xiaoxi Gong

  • Object Recognition and 6D Pose Estimation & Robot Grasping

    Object recognition and pose estimation are of primary importance for tasks such as scene understanding, autonomous car, robotic grasping. However, it has been facing great challenges since severe occlusions, background clutter or large-scale changes. Current methods like CNNs descriptors or Hough Forest, which make use of global or local information to infer the location of object and estimation its pose, still have problems with occlusions as well as time consuming.

    Robot grasping has been widely used in industrial manufacture and family robot. However, most robots grasp object just by using of shape detection or RFID, which has low precision and less robust to bad environment. Our target is to recognize the object and estimate its pose through RGB-D images or point cloud, and then drive the robot grasp it.

    Researcher: Yuanpeng Liu

  • Digital Design and Manufacturing (inspection)

    The rapid development of digital measure technology provides various of measuring means with large range, high accuracy and high efficiency, for aircraft component assembly. At present, both domestic and international aircraft manufacturers accomplish aircraft component assembly with digital measuring equipment, such as laser tracker, indoor GPS, laser radar, total station, etc. Constructing digital measuring field with digital measuring equipment, to position aircraft component accurately, is the foundation and precondition of aircraft component assembly.

    Point clouds data are used to evaluate the quality of manufacturing and assembly parts, For classification and semantic mapping of machining feature and assembly feature, deep learning techniques are used based on CAA platform. We attempt to drive the model using point cloud data to form new model that reflect the real status of the part to analyze the tolerance.

    Researcher: Yan Wang

  • Medical Image Processing

    The most important part of medical image processing is image segmentation. Image segmentation is a procedure for extracting the region of interest(ROI) through an automatic or semi-automatic process within a 2D or 3D image. A major difficulty of medical image segmentation is the high variability of the organs in medical images. We mainly use two approaches for image segmentation: (1) deformable model-based segmentation approaches, and (2) deep convolutional neural network-based segmentation approaches.

    Researcher: Kaicheng Zhang

  • Large-scale LiDAR Point Clouds Analysis

    Terrestrial laser scanners are becoming increasingly important in many fields of mapping applications, providing a great amount of 3D positional information in a fast and precision way for a large scale scene measurement in difficult field conditions. This makes terrestrial high resolution laser scanning a technique that is increasingly being used for 3D scene automatic reconstruction or geodetic deformation measurements of civil technical constructions. In this research, we are focusing on registration, reconstruction and analysis of deformation based on 3D laser scan point clouds.

    Researcher: Cheng Yi